For most of the web's history, a purchase has ended with a human clicking a button. The rise of AI agents that can browse, compare and act on a user's behalf breaks that assumption, and it exposes a gap: the entire checkout system, from card forms to fraud screening, was designed for a person, not a program. Over the past year Stripe has been assembling the plumbing to close that gap, and the pieces now have names: the Agentic Commerce Protocol, Shared Payment Tokens, the Link agent wallet, the Machine Payments Protocol, and an experimental browser capability called WebMCP. This piece explains how those parts fit together, why each one exists, and the problems that remain unsolved.
OpenAI
Google
MetaThe core problem: an agent is not a person
When a person buys something online, a lot of invisible trust is doing work. The merchant trusts that the browser session belongs to a real customer. The card network trusts that the person entering the number holds the card. Fraud systems lean heavily on signals like device fingerprints, typing patterns and browsing history that assume a human on the other end. An autonomous agent breaks all of those assumptions at once. It has no card of its own, it should not be handed the user's raw card number, and it looks, to a fraud model trained on human behavior, exactly like the automated abuse those models are built to stop.
So the task is not simply to let an agent click "buy." It is to build a parallel set of rails where an agent can be given narrow, revocable authority to spend, where the merchant can tell an authorized agent apart from a bot, and where the user stays in control of what gets bought and for how much. That is what the various protocols are each trying to standardize.
The four layers of an agent purchase
It helps to break an agent-driven transaction into the same four steps a person moves through, because each one needed its own solution.
How an agent completes a purchase
The four layers of an agent-driven transaction and the protocols Stripe has proposed or co-authored for each. Each layer replaces a step a human used to do by hand.
The agent finds a product and reads its price and availability.
The agent assembles the order, shipping and tax through a hosted session.
A scoped credential authorizes the charge without exposing card details.
The transaction is screened for fraud before it settles.
WebMCP is an experimental browser capability based on a proposed web standard; support and tool schemas can still change.
Discovery is how the agent learns what is for sale. Stripe's Agentic Commerce Protocol, or ACP, launched in December 2025 and co-developed with OpenAI, defines a hosted endpoint a merchant can expose so agents can read a catalog, prices and availability in a structured way rather than by scraping a web page. Google's side of the same problem is addressed through a Universal Commerce Protocol tied to purchasing inside AI Mode and the Gemini app.
Checkout is where the order, shipping and tax get assembled. Stripe routes this through its existing Checkout Sessions API, which means merchants can reuse the checkout infrastructure they already run rather than building a separate agent path.
Payment is the sensitive step. Rather than exposing a card number to an agent, Stripe uses Shared Payment Tokens, scoped credentials that authorize a specific charge, and a Link agent wallet through which an agent can pay while the user keeps spending approvals and visibility over what is bought. For machine-to-machine and micro-scale payments, Stripe and Tempo co-authored a separate Machine Payments Protocol supporting microtransactions and recurring charges in both stablecoins and fiat.
Risk is the backstop. The transaction still passes through Stripe Radar, the same fraud system used for human payments, adapted to the new signals an agent transaction produces.
The task is not simply to let an agent click buy. It is to build rails where an agent can be given narrow, revocable authority to spend.
On the design problem behind agentic commerce
WebMCP: teaching the page to talk to the agent
The most technically interesting piece is WebMCP, and it addresses a different failure mode. Today, a browser agent that has to complete a checkout it was not explicitly integrated with does so by looking at the rendered page, guessing which pixels are the card field and the submit button, and simulating clicks. That approach, driving the page like a human with a mouse, is brittle: a layout change breaks it, and every step costs the agent tokens to read and reason about the screen.
WebMCP flips the relationship. Instead of the agent scraping the page, the page itself registers a set of structured tools the agent can call directly: read the amount due and available payment methods, select a payment method, fill fields, and, where Stripe controls the submit action, start the payment. Stripe's documentation frames this as improving reliability and efficiency for browser agents, because a tool call with a defined schema is far more robust than inferring intent from a screenshot. The important caveats, stated by Stripe itself, are that WebMCP is an experimental capability built on a proposed web standard, that tool availability and schemas can change, and that agents should discover tools at runtime rather than hard-coding them and fall back to standard browser automation when a needed tool is not present.
Who is actually building on it
The rails only matter if merchants and platforms adopt them, and this is where the picture is still early but real. Stripe has said brands including URBN, the parent of Anthropologie, Free People and Urban Outfitters, along with Etsy, Coach, Kate Spade, Revolve and others, are onboarding to sell through agents. Distribution is coming through commerce platforms such as Wix, WooCommerce, BigCommerce and Squarespace, so a merchant does not have to integrate the protocol by hand. On the demand side, Stripe announced at its Sessions 2026 conference in April a partnership with Meta to enable native checkout inside Facebook ads and one with Google to allow purchases through AI Mode and the Gemini app.
What is notably absent from Stripe's own materials, and worth flagging, is hard adoption data. There are no public figures yet on how many agent-initiated transactions are actually completing, what their fraud and dispute rates look like compared with human checkout, or how often WebMCP succeeds versus falling back to brittle browser automation. The infrastructure is being laid faster than the usage data to judge it, which is normal for a new standard but means claims about agentic commerce's scale should be read with care.
The unsolved problems
Three hard questions sit underneath the plumbing. The first is trust and liability: when an agent buys the wrong thing, or is manipulated into an unwanted purchase through a prompt injection on a malicious product page, who is responsible, and how does a refund flow work when no human approved the order? The second is fraud in reverse: fraud systems are tuned to stop automation, so a legitimate agent has to prove it is authorized, and the mechanisms for that, scoped tokens and wallet approvals, are new and largely untested at scale. The third is concentration: agentic commerce routes buying decisions through whichever agent and whichever payment rail a user relies on, which gives real power to the handful of companies that control those choke points, a dynamic that banks and regulators have already begun to raise as agentic shopping grows.
The takeaway
Agentic commerce is at the stage where the roads are being paved before much traffic is on them. Stripe has done the unglamorous work of turning "an agent buys something for you" into a set of concrete protocols for discovery, checkout, scoped payment and fraud screening, and WebMCP points at a future where web pages expose structured tools to agents rather than forcing them to mimic a mouse. Whether this becomes the default way people shop depends less on the plumbing, which is largely in place, and more on trust, liability and whether the fraud and control mechanisms hold up once real money flows through them at scale. For anyone building agents that act in the world rather than just answer questions, the lesson is the same one that runs through the rest of the AI stack: the value is in an agent that can operate across many services and rails on a user's behalf, under the user's control, rather than being locked to a single provider's ecosystem. Keeping that layer open and portable, so an agent can use whichever tools and models fit the task, is the design principle Metir AI applies to the agents it builds.
Sources:
- Introducing the Agentic Commerce Suite: A complete solution for selling on AI agents | Stripe
- Use WebMCP to complete Stripe payments in a browser | Stripe Documentation
- Everything we announced at Sessions 2026 | Stripe
- Agentic Commerce Protocol | Stripe Documentation
- Stripe Agentic Commerce: Infrastructure for the Agent Economy | Stripe
- WebMCP | Chrome for Developers
Image credits
- Hero and in-body figure: "Stripe Dublin office" by Wikimedia Commons contributor (photo credited on Commons to Conor McCabe Photography), licensed under CC BY-SA 4.0. Source: Wikimedia Commons. Reviewed before publication; clear photograph of a Stripe office building with visible branding, honestly depicting the company. Licence permits editorial and commercial reuse with attribution and share-alike.